Improvement of software reliability modeling predictions by the detection and removal of test outliers

摘要

Modeling the times between failure of software under test is one method for predicting when the software will be ready for release. For various reasons such as poorly written test cases, a chosen software reliability model may over-estimate the mean time to the next failure (MTTF). When a test case shows a longer time to the next defect, it biases the estimation of MTTF and that failure time can be considered to be an outlier. In this paper, order statistics is used to construct a bound such that the probability that the kth largest values (relative to their positions in the ordered series) in the failure dataset will exceed that bound is fixed at a small level of significance. The simulation of contaminated datasets is used in the research to validate the proposed approach. Also, real failure datasets with actual Time To Failure (TTF) data are used to demonstrate the approach.